Can Graph Engineering Create System Intelligence?

Hello community. So great that you are back. Yes, we talk about looping. We talk about graph engineering. We talk about task agent and state as the elements. And we're going to build a new ontology, ontology engineering. The new future of AI. Let's have a look. Official title: From agent loops to system intelligence, INI. Let's have a look. Now, this is the study of today, August 21st, published 2026. Look at those universities. Amazing. You have everything from US University of Illinois campaign here to Say Young University here. Those are, my goodness, absolutely impressive team of ORIs and what they are talking about, about graph engineering in the area of LLM agents. And they say we go now from an individual intelligence in AI to a system intelligence, and they have a beautiful survey, but it is not only a survey, it is also here an outlook and an AI that they present you a way that they think that we should think about system intelligence. So let's have a look. Yeah, you have a GitHub file. Everything is there. You have all the papers with all the links here in the GitHub. So you just go there. All the papers that are relevant for the last three years you can find over there. And the authors tell us here in this paper we systematically review the principle methodologies and applications of graph engineering. So they look back here and say, okay, how does it happen that we came to graph engineering and now graph engineering is the new trend August 21st, 26. And then they say in the future we also have an idea what is going to happen in the future and this is now the view they have now. Let's have a look if this is something you would agree with. The first is the model intelligence, no, a GPT-2 model years ago. Then we moved on to prompt engineering. No, we wrote better prompts. Then we said, "Oh, we have to provide more context to the LLM in the context window." So we optimized the context engineering. Then we found out here the LLM is hallucinating like hell. So we have to build a deterministic harness around it. Harness engineering to limit the hallucination and also provide all the numerical calculation and everything else. And then we said, okay, we have now if you want self-learning, we have to loop engineer this so that the system here is self-learning, self-refining here, and self-improving. And now here, this is exactly where the authors are now, graph engineering, and they say this is here now the hot topic August 2026. And the outlook into the future is ontology engineering. Now you might say, wait a minute, graph engineering as the new sync? I mean, graph engineering we know this. Yeah, well, before we have a definition, look at this. Let's have a definition of an agent. Now please note that here in formula one of the paper, so the first and prime assumption is that an individual AI agent can be abstracted as a loop function over a foundational model, your LLM, the harness configuration that you found beautiful, and the runtime state of an agent here at a particular time t. Now you know this is now a significant limitation because in my definition of an agent I also have these three elements but not a loop function because this would eliminate everything else that is not a loop. So therefore starting here with the basic understanding what is an AI agent, this is a loop function is a very interesting approach to say the least. Now you know I told you we went from loops to graphs a month ago. I showed you here the first video. So four months ago here, here knowledge graph and agent graph massive here with massive skill graph integration two months ago. So you know if you're a subscriber of my channel, graphs, yeah, we work with graph for years now. I mean just think about here the graph layer. The first time I showed you a graph layer was four years ago on this channel. Yeah, when we talked about graph neural network of filtrations and persistent homology and just to remember a year ago Howard presents your new knowledge graph agent, no, for medical AI. So graph engineering is nothing new. So what is now suddenly this idea now? And here you see now this new view of the world. The ORIs tell us that this is the way you have to see AI. And let me be clear, this is the view of the authors of the paper. I do not agree with this but since it is the publication, let's have a look, let's learn together, maybe I'm wrong. So they say beautiful, so we have the model intelligence now, the LLM, the prompt engineering, context engineering, everything that was here our LLM specific. And then they say then we had the individual intelligence of our AI and this was done with harness engineering so we build a harness and we engineer the perfect harness for the perfect model. Yeah, and then we had loop engineering if you want for this recursive self-improvement. Yeah, it now comes the part that is a little bit strange for me. Now they define a system intelligence as the next step in the AI development that has two components. It's the graph engineering and ontology engineering is the future, the further out engineering direction, but now graph engineering. But you know, graph engineering we do for years now before we did even harness engineering. So what is so specific that they define now a system intelligence? It has to be graph engineering and not loop engineering, for example. And you see here I had my first, my first doubts, my first question. I said I'm not really sure that this is a view that I like, that I would agree on. And they beautifully defined here their understanding of graph engineering and they said, you know what, we have more or less three major building blocks now: the task organization, so what must be done and how it should be organized, and we will talk about this, and then the team building and the team agent and the team orchestration on everything, and then the runtime state management. Now if you think about it, yeah, of course I understand the intention of the authors to say we have a crazy hallucinating LLM at the core. So therefore we have to build now here with graph engineering a limitation here of this degree of freedom of the LLM. And how you do it? You say you don't have all degree of freedoms to move in a 10,000 dimensional space. But I give you a graph structure and you are only allowed to move along this particular graph structure. So in this graph structure you can apply to the task organization. This means in my wording the workflow optimization. And I say now in the planning, okay, I'm thinking, I write a plan and then this is a plan. So this plan tells me, okay, I have to use the tool A and then use the tool B and then have to do this. And you see I'm heavily restricting the freedom of reasoning of coming up with a good solution for this particular agentic system because I say, listen, you are only allowed to follow here my predefined mapping, here my graph structure. That reduces here the complete path complexity for the system in total. And the same is done here for the agent team communication. How do you have a planner, a coder, a researcher, reviewer, agent, you know, all of this becomes a graph? Yeah, no problem. We did this three years ago. And then we have runtime state manager. Exactly the same logic. So I absolutely understand that they say, hey, graph engineering is the new thing where we reduce the complexity of a freely hallucinating LLM. But I think this is the state-of-the-art three years ago. And then it's interesting because they say, okay, so why do we have now to go from individual intelligence here to a system intelligence? And they have here paragraph 0.5 limitation of individual intelligence since individual intelligence is typically organized around a single agent. And now it made click in my brain. I said, okay, now I understand why they have chosen this. So for beginner in AI to be real simple, you know, and not overload this, they say listen, an individual agent or individual intelligence is a single agent and here we have a single harness and here a single loop engineer. And then if I want to go to

new future of AI. Let's have a look. Official title from agent loops to system intelligence ini. Let's have a look. Now this is the study of today August 21st published 2026. Look at those universities. Amazing. You have everything from us University of Illinois campaign here to say young

university here. Those are my goodness absolutely impressive team of oris and what they are talking about about graph engineering in the area of LLM agents and they say we go now from an individual intelligence in EI to a

system intelligence and they have a beautiful survey but it is not only a survey it is also here an outlook and an a that they present you a way that they think that we should think about system intelligence.

So let's have a look. Yeah, you have a GitHub file. Everything is there. You have all the papers with all the links here in the GitHub. So you just go there all the papers that are relevant for the last three years you can find over

there. And the authors tell us here in this paper we systematically review the principle methodologies and applications of graph engineering. So they look back here and say okay how does it happen that we came to graph engineering and

now graph engineering is the new trend August 21st 26 and then they say in the future we also have an idea what is going to happen in the future and this is now the view they have now let's have a look if this is something you would

agree with the first is the model intelligence no a GPD2 model years ago then we moved on to prompt engineering no We wrote better prompts. Then we said, "Oh, we have to provide more context to the LLM in the context

window." So we optimized the context engineering. Then we found out here the LLM is hallucinating like hell. So we have to build a deterministic harness around it. Harness engineering to limit the the hallucination and also provide all the numerical calculation and everything else. And then we said okay

we have now if you want self-arning we have to loop engineer this so that the system here is self-arning self-refining here and selfimproving and now here this is exactly where the authors are now graph engineering and they say this is here now the hot topic August 2026 and the outlook into the

future is ontology engineering now you might say wait a minute graph engineering as the new sync I mean graph engineering we know this yeah well before we have a definition look at this let's have a definition of an agent now please note that here in formula one of the paper so the first

and prime assumption is that an individual EI agent can be abstracted as a loop function over a foundational model your llm the harness configuration that you found beautiful and the runtime state of an agent here at a particular

time t. Now you know this is now a significant limitation because in my definition of an agent I also have these three elements but not a loop function because this would eliminate everything else that is not a loop. So therefore starting here with the basic

understanding what is an EI agent this is a loop function is a very interesting approach to say the least. Now you know I told you we went from loops to graphs a month ago. I showed you here the first video. So four months

ago here here knowledge graph and agent graph massive here with massive skill graph integration two months ago. So you know if you're a subscriber of my channel graphs yeah we work with graph for years now I mean just think about

here the the graph layer. The first time I showed you a graph layer was four years ago on this channel. Yeah, when we talked about graph neural network of filtrations and persistent homology and just to remember a year ago Howard

presents your new knowledge graph agent no for medical AI. So graph engineering is nothing new. So what is now suddenly this idea now and here you see now this new view of the world the orers tell us that this is the way you have to see AI

and let me be clear this is the view of the authors of the paper I do not agree with this but since it is the publication let's have a look let's learn together maybe I'm wrong so they say beautiful so we have the model

intelligence now the LLM the prompt engineering context engineering everything that was here our L&LM specific and then they say then we had the individual intelligence of our AI and this was done with harness engineering so we build a harness and we engineer the perfect harness for the

perfect model yeah and then we had loop engineering if you want for this recursive self-improvement yeah it now comes the part that is a little bit strange for me now they define a system intelligence as the next step in the AI

development that has two components It's the graph engineering and ontology engineering is the future the further out engineering direction but now graph engineering but you know graph engineering we do for years now before we did even horn engineering so what is so specific that they define now a

system intelligence it has to be graph engineering and not loop engineering for example and you see here I had my first my first doubts my first question I said I'm not really sure that this is a view that I like that I would agree on

and they beautifully defined here their understanding of graph engineering and they said you know what we have more or less three major building blocks now the task organization so what must be done and how it should be organized and we will talk about this and then the team building and the team

agent and the team orchestration on everything and then the runtime state management Now if you think about it, yeah of course I understand the intention of the authors to say we have a crazy hallucinating LLM at the core. So therefore we have to build now here with

graph engineering a limitation here of this degree of freedom of the LLM. And how you do it? You say you don't have all degree of freedoms to move in a 10,00 dimensional space. But I give you a graph structure and you are only

allowed to move along this particular graph structure. So in this graph structure you can apply to the task organization. This means in my wording the workflow optimization and I say now in the planning okay I'm thinking I write a plan and then this is a plan. So this plan tells me okay I have to use

the tool A and then use the tool B and then have to do this and you see I'm heavily restricting the freedom of reasoning of coming up with a good solution for this particular agentic system because I say listen you are only

allowed to follow here my predefined mapping here my graph structure that reduced here the complete path complexity for the system in total and the same is done here for the agent team comm communication. How do you have a

planner, a coder, a researcher, reviewer, agent, you know, all of this becomes a graph? Yeah, no problem. We did this three years ago. And then we have runtime state manager. Exactly the same logic. So I absolutely understand that they say, hey, graph engineering is the new thing where we reduce the

complexity of a freely hallucinating LLM. But I think this is the state-of-the-art 3 years ago. And then then it's interesting because they say okay so why we have now to go from individual intelligence here to a system intelligence and they have here paragraph 0.5 limitation of ind

individual intelligence since individual intelligence is typically organized around a single agent and now it made click in my brain. I said okay now I understand why they have chosen this. So for beginner in AI to be real simple you

know and not overload this they say listen an individual agent or individual intelligence is a single agent and here we have a single harness and here a single loop engineer and then if I want to go to one plus

this means two and more agents this is then a system intelligence and then for the coordination of multiple agents I need graph of engineering. So you see suddenly I have now the right glasses the right perspective and suddenly I

understand what the authors wanted to show us here if you're new to AI if you are a student and you say I want to have the basic understanding they want to help you understand some basic things but honestly between you and me here and

you as a subscriber of my YouTube channel you know this is incorrect this is mathematically incorrect and this is incorrect if we apply the logic of computer science so this is just the a very first approximation to help

somebody neutrii to come up with some construct some sort processes but we know operative this is incorrect this is not what we do anyway let's let's see you know sometimes it helps me to read something and then reading this paper I thought multiple times wait a minute this is

nonsense but it helped me to clarify precisely ly reading these sentences that I disagree with this. So I learned from reading the paper that I do not agree with this. But you know what? Do not trust my judgment. You you have to

read the paper. You have to go there and you have to develop your own knowledge and your own view how you want to see this. And maybe you say no I agree with them and this is absolutely fine. This is here call it whatever critical

thinking or thinking for yourself or whatever you have to get a feeling here. If somebody here I mean the top tier university of this planet they present to you here official paper and then I come and I tell you I do not agree with

this. You might say hey what is happening? Are you going crazy? No just look at the topic. So they are graph engineering as I told you here the task organization and they say task organization structures. Now the objective into explicit subtask and when you read this you know exactly what is

happening. You know that the order say hey the eye is not able to handle the objective of of the task of the query. Therefore the I has to subdivide the higher complexity into multiple lower complexities multiple subtask and then

we reduce the subtask to sub subtask and then we break it down to even simpler task and sometimes the AI will be able to solve this task and this is why we have the task organization now subdivided into subtask here and into

workflow structures that are now have a graph structure. So this means you define now a deterministic workflow structure for a probabilistic LLM and you understand immediately why the oring this is here a good representation.

Now the second is here exactly the agent coordination defines the team topology. How many agent do you need? Do you go here with a boss agent? No it defines here how it routes the communication among the agents and so on. You go here

for a orchestration agent or you go for a swarm. All of this and then you have the runtime. Yeah, you have the same detects and localizes anomalies.

Supports recovery and structural updates. But you see all of them have together you want here to limit the available search base the mathematical space here to reduced graph representation.

Now if you go on in chapter four they talk about graph engineering now and why they need graph engineering from individual intelligence to system intelligence. No and they tell us now and now graph provide a natural structure for modeling the system level relationships.

And of course this is true but you know what graph also are extreme helpful to organize and model the single agent relationships and especially [snorts] if the agent has llm and multiple harness components. I do this in a graph

structure. So therefore the way I implement the eye I say this is not the complete correct way I would say that this sentence is true.

And I give you here more details about the goal decomposition. But it's always the same. Now divide divide divide divide divide. Make it simply simply simpler. And yeah and then confirm. Same with workflow optimization.

Log par analyzer sandbox runner patch generator test runner fixed versus. Yes. Yes. Yes. Yes. And guess what it same happens here with the team building.

The same is happening for the runtime. So honestly now the paper tells you this is why we need graph engineering. This is it to go to multiple agents. And then the paper tells you and this is even even more interesting because I do not

agree. You will be surprised with point number seven that the future of AI is ontology engineering. So let's have a look at this. But please read the paper. Do not trust my statement. You have to have your own

opinion. They go in paragraph 5.2 two and they say self- evvolving graph system now and say yeah let's have a look I just had my last videos now about self- evvolving multistructure system with graph and with looping and everything now and they provide us here and this is a screenshot

here but just look at this no aflow I did my video on aflow exactly more than a year ago so okay this gives you an indication where they are with their understanding ing of self- evvolving graph systems.

But let's come here to the future. So this yeah paragraph six here in the paper is future direction and they say it is more or less here focus on ontology engineering for the next generation system intelligence. So they

go again with their definition of system intelligence. They define the next generation of this particular term and they tell us and you know what is the AI technology for this? It will be ontology engineering.

So ontology engineering is I thought not what I understand ontology engineering as a theoretical physicist but what they understand it. So here we go again they have some beautiful images. So this is why I can show you so much

screenshots from the original paper again the model intelligence the individual intelligence then in their view the moment you go from one agent to multi- agents you have system intelligence and then you need a graph which I have already my graph implementation here and I mean just to

tell you know this is here a neural network architecture can also be understood as yeah I'm not [clears throat] and here you have the future ontology engineer this is our future path according to the orus and then I looked a little bit closer on their sources and I saw that for the

ontology engineer I mean the first what I noticed is Palanteerontology now I don't know if you're familiar Palentia technology a US uh software company primarily here specialized for governments secret services military yeah police military you Got the idea now. Palunteer foundry for here. Yeah.

Economic system for a civil institution here. So yeah, but now let's talk about it. Palunteer is absolutely necessary that they have an ontology because if they go here for let's say military or secret service or whatever they go, they

need here exactly a formal structured framework. Who are you? What is your name? what is your occupation, where are you living, what is your uh tax identification number or whatever you have, social security number or whatever. So they need a formal structured framework. So whatever you

are here in the surveillance business with AI, I understand that ontology is absolutely necessary. You have to have a formal structured framework used to represent knowledge by defining here a set of concepts within a particular domain and the relationships between them.

Now if you think about what is a relationship, how can I present a relationship? Maybe you have a node and an agent. [clears throat] Okay. So I went now to Google and here the core components of a data ontology here are

classes. This simply means categories of types of objects like company, employee. Got it? Then we have our attributes and properties. No example company has revenue. Got it? Relationship rule that define how classes connect to one another. Let's say an employee works for a particular company and then you have

axioms no logical constraints. So I understand that if you are into the surveillance business and you monitor I don't know all the criminals in the world you have to know exactly where they are what they doing what is their

education what other people they know where they have been working for what they have done on the day 312 of the year great but do we need ontology charts for the future of AI if we are not into this business

have a look at this paper this was here this is already version three here from June 2026 and this is here about culturally aligned LLM through ontology guided multi-agent reasoning and they say listen we have to understand here

what is specific about nations no we have to understand here the cultural differences everybody is has a different cultural background so if we want to provide or want to sell a product to somebody in Africa or somebody in Asia

or somebody in in Europe we have to understand their cultural background and if we lack demographic information. So okay you might say okay here too we need ontology or if you go here a study by University of Cambridge ontology to

tools compilation for executable semantic constraint enforcement in LLM agency from February 26 you have now particular knowledge graph representation let's say in chemistry or in biochemistry and then you need to have for each molecule or each possible combination of molecular level exactly

here the ontology um instances that you know Hey, this has the function has synthesis step width or he has a chemical input or has a chemical output.

Yeah, absolutely this is the future if you go with this. But for general LLM agents, the graph engineering is a good old friend of us. Rough engineering is not something that is suddenly new and suddenly beautiful.

So are they all incorrect? Is the view of all of this institution here suddenly incorrect? Well, it is just from the authors of these papers from this institution their general view of AI and AI is an extreme complexity. But I know having read specific individual highly specific

paper from each of these institution in the last four years I know exactly that they also work here on understanding of AI that is closer to the understanding of AI you find in my videos. But of course if you want to have here a general paper for the very

educated people now okay you might have some simplifications. So yeah, you have to be extremely careful if you read this paper to really look at the terms, look at the definition, understand how they are building from what how they start building up their theory because maybe

it is not what I go with. Let's come now to the official critique. Yes. And I dare to critique those institutions. I love this institution. I love all the orders of this institution. But here for a scientific purpose now to

the critique section. Now if we if we really have to divide as the authors of the paper do between loop engineering and graph engineering I have a simple view they describe different mathematical dimensions because loop goes more or less with time and graph engineering goes with relational

structure organization and real system simply require both. So therefore separating here the loop engineering into the individual part and separating the graph engineering into the multiple system part is nonsense in my view.

A loop describes a temporal dynamics. We go from x at t to x at t +1. A graph describes a relational structure where we have a relation between here the nodes and we have edges, hyperraphs, hyper structures, whatever you want. But

notice that normally these are orthogonal structures. So it makes no sense to integrate a loop only in an individual intelligence and not have a loop in uh multi- aent intelligence. I cannot agree with the orders in this particular point because I see loop and graph as at least autoonal and I need

both of them. I also think critique number three there's a boundary problem. Think about if one agent harness can spawn multiple sub agent, manage a task deck and maintain some shared state over the time. Is it still a single agent or is it already an agent system? This

presentation by the or does not make sense for me in an operational way. Now I don't have to tell you since you are a subscriber of my channel that every workflow, every program, every state machine or any organization can be

represented as a graph. It just depends you on the particular graph representation that you choose your particular representation your particular mathematical background how you want to view it what mathematical formulas you want to apply it and yeah you have graph native operation that can

maybe outperform other operation but you have to validate this and I operate now more with loops than with graph structures but it simply depends on the complexity of the problem. Another critique would be just having multiple agents like presented in a paper does not guarantee

independent let's call it let's go extreme super intelligence. No just because you have multiple agent does not that are now configured in a graph does not mean that this system has a better performance. This would be a critical

mistake. Just because you have a coder agent and a reviewer agent here, two instances may create some organizational separation without some genuine epistemic independence at all and it might even harm the result of the complete system. So I do not agree with the orders on

this. Now what are the limits? If we go now to point 7, the ontology, an ontology can define absolutely evidence, completion, authorized action. Absolutely.

But what it cannot establish is an observation. Is this factually correct or not? That the relation is causal or that the chosen goals are desirable or whatever the machine comes up with. Yeah. So I think or I've seen that a perfect consistent ontology beautiful in

theoretical physics and mathematics and believe me much more complicated than you can imagine can encode a perfectly wrong theory.

So just because you have a complete ontology established from your data complexity in your particular domain does not mean that you know the ground truth or that you know what is factually correct. It can go and just yeah

dissipate here fully in the wrong theory. Another point of critique since we're here a fixed ontology that is done in the planning can obstruct any scientific discovery. So if the ontology already defines all the admissible entities and the relation and everything in your

structure, the system can think about it detect only the known unknowns. So if the system knows I only have information about nine out of 10 points and I know there's the 10th point that I have no information about. This is known

unknowns. But what if there's 11 and2 to your data that you have no idea exist?

What if this let's say this criminal that you are super revising here with whatever technology has some complete other relation to other people you're not monitoring that would provide a complete different view on the actions on this particular person.

So whenever you go with fixed topologies just be just be absolutely careful. Graph maintenance go back here to the point of graph may become its own bottleneck. We have seen it in my videos. Now and just to be clear the

paper does not establish when the infrastructure procedures here provide a net benefit over a simpler agent loop structure. So just doing a literature survey is maybe not enough.

So therefore let's come to my conclusion. I have a very simple view on AI if you ask me. I say a graph simply tells me hey what is connected? An ontology says what this connection means in a particular domain since we have defined what it is. And then the runtime

determines what the connection does in reality. If I activate it, if I simulate it in a sandbox, I really have it the inference run or whatever. So you see the topology, the simplicity of a graph ontology and runtime. I see it in a

simple view. What can an ontology not do? And I think an ontology cannot guarantee factual correctness. An ontology cannot establish ground truth. An ontology cannot ground concepts in the physical world. An ontology cannot prove that a metric measures anything like similar to

intelligence. In my view, ontology engineering is nice. It's absolutely necessary for some task. But understand ontology engineering is if you want it a little bit more to the point semantic governance but it is not an autonomous source of the truth.

So therefore if I look now in the future absolutely in contradiction to your paper to the authors of this paper I would say system intelligence if I today would just have to define it system intelligence is a collection of local

loops organized by evolving dynamic graph structures constrained by domain specific shared ontologies and grounded by external evidence.

This is quite an interesting video and I thought about should I do it and I decided yes because maybe I am wrong and maybe you read the paper and you decide to go with all of these authors and all of this renowned university and

institution and whatever. But I thought it is also sometimes helpful to show you if you look at this paper you sometimes have based on your own knowledge feelings that tell you no this is not something I would agree with. No. And if

there are 100 universities that claim this I know from my experience from my logic from my understanding of mathematics and AI this is a contradictionary statement that is not valid. So therefore whatever you read on scientific literature be careful activate your own neural network and you

can enjoy so many papers so much literature it is simply amazing and you know what reading this particular paper also I do not agree with it especially in point 6 and 7 I thought it was really helpful interesting and I would

recommend reading the paper wherever you suddenly start to hear this little voice in the in your ad that says, "Hey, wait a minute. This this is not correct. No, this is not the way I see it. No, an agent is not defined over only a loop

structure. No, this is not because this gives you a beautiful opportunity to learn, to criticize, produce a video and maybe you will be heavily critiqued in the comments of this video. So, let's see. Let's do this experiment together.

Would be great to see you in my next video.